McCoy and colleagues predicted that, because language models are trained to predict the next word, their accuracy should depend on how probable the task, input and output are, even where probability should not matter; GPT-4 decoded a simple cipher with 51% accuracy when the output was a high-probability word sequence and 13% when it was low-probability.

SocioLens project,

Claim LD01-SL-C18 (v1.0)

Research opportunities this claim motivates

What this claim leaves open for research. Each is a hypothesis about a gap, not a finding.

Do the failure patterns of 2023–2024 language models persist in reasoning models?Accuracy tracks output probability in GPT-4 (51% against 13% on a cipher).

Provenance and status

Claim type: Estimate and Interpretation
Publication status: Eligible for publication, with qualifiers: preprint
Verification state: Checked against the primary source; the source is a preprint. Who checked this claim is not recorded for it; claim checks so far were AI-led, and no person has checked claims against their sources.
Does not license: A claim that models cannot do these tasks, or that they lack understanding; the authors say the teleological view can be incomplete
Limitations: Read in arXiv version 1 (24 September 2023), which covers GPT-3.5 and GPT-4 with basic prompting; the PNAS version (2024) was not compared; frontier models have changed since
Source (as cited): McCoy, Yao, Friedman, Hardy & Griffiths, “Embers of Autoregression: Understanding Large Language Models Through the Problem They are Trained to Solve,” arXiv:2309.13638, 2023
Source type: Preprint (recorded as “Preprint (a revised version was published in PNAS)”)
Institutional position: Independent academic (Princeton University)
Maturity of the evidence: Emerging evidence
Recorded date: 16 October 2026
Currentness / correction: This is the current version. No correction or concern is recorded on its sources.

Sources this version cites

McCoy, Yao, Friedman, Hardy & Griffiths, “Embers of Autoregression: Understanding Large Language Models Through the Problem They are Trained to Solve,” arXiv:2309.13638, 2023 (Preprint)
Read the original

Version

This is the current version, v1.0. Earlier versions stay available, and their records never change.

Versions: v1.0 · current

Raw record: every recorded field, with its coded values

All 25 canonical Claim fields, verbatim, with the coded values derived from them. The plain-language labels above are read from these.

Claim type (coded): ESTIMATE + INTERPRETATION (as printed: “ESTIMATE / INTERPRETATION”)
Publication status (coded): ELIGIBLE (PREPRINT) (as printed: “ELIGIBLE WITH PREPRINT QUALIFIER”)
Verification state (coded): CONTENT_VERIFIED_PRIMARY (PREPRINT)
Source type (coded): PREPRINT
Maturity (coded): EMERGING

#	Canonical field	Canonical value (verbatim)	Normalized (Claim Value Vocabulary v1.0.2)
1	CLAIM ID	LD01-SL-C18	—
2	PUBLIC-FACING CLAIM	McCoy and colleagues predicted that, because language models are trained to predict the next word, their accuracy should depend on how probable the task, input and output are, even where probability should not matter; GPT-4 decoded a simple cipher with 51% accuracy when the output was a high-probability word sequence and 13% when it was low-probability.	—
3	CLAIM TYPE	ESTIMATE / INTERPRETATION	base_types: ESTIMATE, INTERPRETATION
4	RESEARCH STRAND	B — Evidence that could tell the accounts apart	—
5	MATURITY	EMERGING	states: EMERGING
6	SOURCE	McCoy, Yao, Friedman, Hardy & Griffiths, “Embers of Autoregression: Understanding Large Language Models Through the Problem They are Trained to Solve,” arXiv:2309.13638, 2023	source_ids: SRC-SL-MCCOY-ET-AL-EMBERS-ARXIV-2309-13638
7	SOURCE TYPE	Preprint (a revised version was published in PNAS)	classes: PREPRINT
8	INSTITUTIONAL POSITION	Independent academic (Princeton University)	—
9	SCOPE	Eleven deterministic tasks in English; models available in 2023	—
10	DESIGN	Experiments varying the probability of tasks, inputs and outputs	—
11	SAMPLE / DATASET	Eleven tasks	—
12	MODEL / TOOL / VERSION	GPT-3.5 and GPT-4 (the PNAS version adds models including Claude 3)	—
13	COMPARATOR	High- and low-probability versions of the same task	—
14	MAIN RESULT	Accuracy tracks probability: 51% against 13% on the cipher task; the authors state comparative claims (“X is harder than Y”), not impossibility claims	—
15	INFERENTIAL TARGET	Whether the training objective leaves predictable traces in behaviour	—
16	DOES NOT LICENSE	A claim that models cannot do these tasks, or that they lack understanding; the authors say the teleological view can be incomplete	—
17	VERIFICATION STATE	CONTENT VERIFIED PREPRINT	state: CONTENT_VERIFIED_PRIMARY · qualifiers: PREPRINT · verified_by: UNKNOWN
18	LIMITATIONS	Read in arXiv version 1 (24 September 2023), which covers GPT-3.5 and GPT-4 with basic prompting; the PNAS version (2024) was not compared; frontier models have changed since	—
19	CONTRADICTORY / SUPPORTING SOURCES	Other limits of the same kind: LD01-SL-C19, LD01-SL-C20	—
20	GEOGRAPHY / LANGUAGE	English task set	—
21	DISTRIBUTIONAL LEDGER	Not analysed in this claim	—
22	R01 / FIELD-MAP LINKS	LD-01, T4, D3	—
23	WHAT WOULD CHANGE THIS CLAIM?	Newer models whose accuracy no longer depends on output probability	—
24	PUBLICATION STATUS	ELIGIBLE WITH PREPRINT QUALIFIER	state: ELIGIBLE · qualifiers: PREPRINT
25	RECORDED DATE	2026-10-16	—

Field 22 references resolved against Field Map v0.2: LD-01 (LIVING_DEBATE); T4 (TENSION); D3 (DOMAIN)

The sources as recorded, with their record IDs and coded source types:

Pinned Sources (field 6, version-specific)

“McCoy, Yao, Friedman, Hardy & Griffiths, “Embers of Autoregression: Understanding Large Language Models Through the Problem They are Trained to Solve,” arXiv:2309.13638, 2023” → McCoy et al. (2023), “Embers of Autoregression…” (SRC-SL-MCCOY-ET-AL-EMBERS-ARXIV-2309-13638--v1.0) — McCoy, Yao, Friedman, Hardy & Griffiths, “Embers of Autoregression: Understanding Large Language Models Through the Problem They are Trained to Solve,” arXiv:2309.13638, 2023; class PREPRINT (explicit public provenance pin)

Cited by Living Debate versions

What kind of thing is generative AI, and does the answer predict anything? (v0.1) LD-01

Cite this page

SocioLens project. (2026, October 16). McCoy and colleagues predicted that, because language models are trained to predict the next word, their accuracy should depend on how probable the task, input and output are, even where probability should not matter; GPT-4 decoded a simple cipher with 51% accuracy when the output was a high-probability word sequence and 13% when it was low-probability [Claim LD01-SL-C18, version v1.0]. SocioLens. https://sociolens-web.pages.dev/claims/ld01-sl-c18/v1.0/

@misc{sociolens_ld01_sl_c18_v1_0,
  author = {{SocioLens project}},
  title = {McCoy and colleagues predicted that, because language models are trained to predict the next word, their accuracy should depend on how probable the task, input and output are, even where probability should not matter; GPT-4 decoded a simple cipher with 51\% accuracy when the output was a high-probability word sequence and 13\% when it was low-probability.},
  year = {2026},
  date = {2026-10-16},
  howpublished = {\url{https://sociolens-web.pages.dev/claims/ld01-sl-c18/v1.0/}},
  note = {Claim LD01-SL-C18, version v1.0}
}

The link points to this version, whose record never changes. For a finding, cite the original study as well: SocioLens summarizes research and does not replace it.

Research-of-record provenance

Materialized by sociolens-materializer 0.6.0 at 2026-09-28T20:00:00.000Z; normalization contract SocioLens — Claim Value Vocabulary v1.0.2 — Pre-Spike Materialization Contract (v1.0.2). Content seal SHA-256 223bc3555dceffe8e0d57dc94addd557e1768135f9000d2e710586d1e6adf1ca.

Repository-authored inputs (not Drive exports)

SocioLens claim — LD01-SL-C18 v1.0 (v1.0) — written in the SocioLens repository by Project Lead (Claude, AI) on 2026-10-16; no Drive identity exists. Authority: AGENTS.md "Authority" (the Project Lead decides research content within the research-integrity rules) and the Owner's choice in the SocioLens project chat on 16 October 2026 to build the LD-01 debate; a new Claim, first version, written from the project's LD-01 source notes, checked against full texts, official web pages or official abstracts. Block “ENTRY 18 — LD01-SL-C18” SHA-256 47f71682e57f155bf74d1d7a157663abd44c0a428d2b1124827a38795fbe5a75; fixture fixtures/authored/claims/LD01-SL-C18--v1.0.

Builder decisions (not Drive data)

VERSION = v1.0 — basis: Claim version taken from the repository-authored claim fixture version (SocioLens claim — LD01-SL-C18 v1.0); the first version of a new Claim of the LD-01 debate
CURRENT_POINTER = true — basis: repository-authored claim "SocioLens claim — LD01-SL-C18 v1.0" (Project Lead (Claude, AI), 2026-10-16); basis: AGENTS.md "Authority" (the Project Lead decides research content within the research-integrity rules) and the Owner's choice in the SocioLens project chat on 16 October 2026 to build the LD-01 debate; a new Claim, first version, written from the project's LD-01 source notes, checked against full texts, official web pages or official abstracts
FIELD_6_SOURCE_REFERENCE = SRC-SL-MCCOY-ET-AL-EMBERS-ARXIV-2309-13638--v1.0 — basis: field 6 of ENTRY 18 — LD01-SL-C18 cites "McCoy, Yao, Friedman, Hardy & Griffiths, “Embers of Autoregression: Understanding Large Language Models Through the Problem They are Trained to Solve,” arXiv:2309.13638, 2023"; the only field-6 segment of this ClaimVersion's repository-authored ledger block.
SOURCE_PROVENANCE_PIN = SRC-SL-MCCOY-ET-AL-EMBERS-ARXIV-2309-13638--v1.0 — basis: explicit version-pinned public provenance of this ClaimVersion (Amendment v1.2 §6): the pinned field-6 Source record "McCoy, Yao, Friedman, Hardy & Griffiths, “Embers of Autoregression: Understanding Large Language Models Through the Problem They are Trained to Solve,” arXiv:2309.13638, 2023".